Ilya Sutskever

Ilya Sutskever is a neural-network researcher and OpenAI cofounder who jointly authored major work on image classification, sequence-to-sequence learning and adversarial examples. OpenAI announced his departure as chief scientist in May 2024; in July 2025, he announced that he was formally CEO of Safe Superintelligence Inc. (SSI), with its technical team continuing to report to him. His later public research agenda emphasizes reliable generalization: models learning in ways that carry beyond their training tasks.[2][3][4][5][8][10][11]

Before those leadership roles, Sutskever worked on concrete problems in training neural networks and understanding their failures. His earlier research provides technical context for his later concern with model reliability.[3][4][5][11]

Sutskever studied in the University of Toronto’s machine-learning group with Geoffrey Hinton, then held a Stanford postdoctoral position in Andrew Ng’s group. He subsequently cofounded DNNresearch and spent three years as a research scientist at Google Brain.[2]

With Alex Krizhevsky and Geoffrey E. Hinton, he authored the 2012 paper ImageNet Classification with Deep Convolutional Neural Networks. Their 60-million-parameter network classified images into 1,000 categories. The team used non-saturating neurons and an efficient GPU implementation of convolutional networks to speed training, alongside regularization to reduce overfitting. On the LSVRC-2010 test data, they reported top-1 and top-5 error rates of 39.7% and 18.9%, respectively.[3]

Sutskever also coauthored Intriguing properties of neural networks, first submitted in December 2013, with Christian Szegedy, Wojciech Zaremba, Joan Bruna, Dumitru Erhan, Ian Goodfellow and Rob Fergus. The team found imperceptible image perturbations by maximizing a network’s prediction error, causing it to misclassify the image. They reported that the same perturbation could also fool a different network trained on another subset of the dataset—showing that the failure could transfer between models.[5]

In 2014, Sutskever, Oriol Vinyals and Quoc V. Le introduced an end-to-end method in Sequence to Sequence Learning with Neural Networks. One multilayer Long Short-Term Memory network (LSTM) encoded an input sequence into a fixed-dimensional vector; another deep LSTM decoded the target sequence from that vector. A consequential training decision was to reverse the source sentence’s word order while leaving the target unchanged. The authors explained that this introduced shorter source-to-target dependencies, easing optimization and markedly improving performance.[4]

On WMT’14 English-to-French translation, the authors reported a BLEU score of 34.8 for LSTM-generated translations across the entire test set, with penalties for out-of-vocabulary words, compared with 33.3 for a phrase-based statistical machine-translation system. In a separate setup, using the LSTM to rerank 1,000 hypotheses from that system raised its score to 36.5.[4]

Sutskever cofounded OpenAI, whose December 11, 2015 founding announcement named him research director. He later served as chief scientist. On May 14, 2024, OpenAI announced his departure and named Jakub Pachocki as his successor.[2][6][8]

Stanislas Polu recalled Sutskever’s involvement in OpenAI’s reasoning team in a November 2024 interview. The team had its own manager, but Sutskever was an active advisor who coached Polu as an engineer becoming a researcher. Polu remembered him moving between teams to communicate a shared direction, and was surprised by how much time he devoted to communicating that vision and getting teams to work together.[12]

In a July 3, 2025 update, Sutskever announced that he was formally SSI’s CEO, Daniel Levy was president and the technical team continued to report to him. SSI’s stated mission is “one goal and one product: a safe superintelligence.” The company says it approaches safety and capabilities together as technical problems, with a business model intended to insulate research from short-term commercial pressures.[10][9]

In his November 25, 2025 interview with Dwarkesh Patel, Sutskever described SSI as an attempt to investigate promising ideas about generalization, whose correctness remained a research question. He also said Meta had offered to acquire SSI and that he had declined.[11]

Sutskever left SSI’s plan to pursue superintelligence directly tentative. Longer timelines could change it, he said, as could the value of letting the public experience powerful AI. He described a change in his own thinking: “I now place more importance on AI being deployed incrementally and in advance.” He said this might affect SSI’s plans, while envisioning gradual release even under a direct pursuit of superintelligence.[11]

In the same interview, Sutskever argued that AI was returning to an era of research with large computers. He acknowledged that more scale would change results, but questioned whether another hundredfold increase alone would transform everything. His central concern was that models “generalize dramatically worse than people.” To illustrate the gap between impressive evaluations and practical reliability, he described a coding assistant that fixes one bug, introduces another and then restores the first when asked to fix the second.[11]

Sutskever proposed value functions as a way to make reinforcement learning more efficient. Instead of waiting until a long task ends to obtain a training signal, an agent could recognize that an intermediate direction is unpromising and learn from that earlier decision. He qualified the proposal: the same capabilities could be obtained without value functions, but more slowly.[11]

His hypothetical superintelligence was also a continual learner, rather than a finished system already possessing every job-specific skill. He imagined a highly capable student acquiring knowledge through deployment and trial and error. On alignment, he asked whether fragile learning and optimization of human values were themselves instances of unreliable generalization; the effect of substantially better generalization remained an open question.[11]

Sutskever described his personal research taste as guided by “beauty, simplicity, elegance, correct inspiration from the brain.” That top-down conviction, he explained, helps him decide whether a disappointing experiment calls for more debugging or abandoning a direction: an apparently negative result may come from a bug even when the underlying idea is sound.[11]